Tail risk in aviation decisions

Before a flight, a lot of pilots fill out some version of a Flight Risk Assessment Tool. You score a handful of factors — weather, how current you are, the route, terrain, the aircraft, time of day — add them up, and land in a green, yellow, or red band. It is a genuinely good idea. It replaces a vague feeling with a number, it forces you to look at each factor on purpose, and it makes the decision consistent from one day to the next. Structure beats vibes.

But the number it produces is an average, and a go/no-go decision is not a bet on the average day. It's a bet on the worst plausible version of the day. Those are different questions, and the gap between them is where the interesting failures live.

Two flights, one score

Imagine two flights that come out to the same middling risk score. The first is a steady flight with no single alarming factor — everything is a little elevated, nothing is scary. The second gets to the same total by averaging a benign cabin with a genuinely volatile sky: a front that might pass north, or might not. Same average. Wildly different tails. The first flight's bad day is a little worse than its typical day. The second flight's bad day is a different category of event.

A single number hides the shape of the distribution. And in risk, the shape is the whole point.

Two things make the standard matrix blur this. First, it collapses a distribution to a point — you never see the spread. Second, the 5×5 grid is coarse and discontinuous: nudge one input and the score can jump a whole band, or not move at all, for reasons that have more to do with rounding than with reality. You can feel both effects directly in the risk explorer — flip between the matrix and the smoother models and watch the same inputs tell different stories.

Borrowing a better question from finance

Quantitative finance ran into exactly this problem and mostly moved past it. For years the standard risk number was Value at Risk — "we're 95% confident we won't lose more than X." The trouble is that VaR says nothing about the other 5%. It draws a line and refuses to look past it, which is precisely the region you care about. So the field increasingly uses Conditional Value at Risk (also called expected shortfall): not "how bad is the line," but "how bad is it on average when things go past the line." It's the mean of the worst outcomes rather than a threshold you hope not to cross.

Bring that lens to a flight and the volatile-weather case stops hiding. Its mean might match the steady flight, but its CVaR is much higher, because the worst tenth of its simulated days are genuinely dangerous. The number now reflects the question the pilot was actually asking.

What a decision tool should actually show

None of this argues for a more complicated black box. It argues for the opposite — showing more of the work, not less. A risk tool worth trusting should, to my mind, do four things:

• Decompose the score. Show each factor's contribution, so the pilot can see whether the risk is broad or concentrated in one place they might mitigate.
• Show the spread, not just the mean. An uncertainty band is more honest than a single point, and it's cheap to compute.
• Surface the tail. Put the worst-case average next to the average, because that comparison is the actual decision.
• Keep the human in the loop. The tool's job is to make the risk legible, not to make the call. The pilot makes the call.

That's the thesis behind FlightReady, and it's the same principle I keep coming back to elsewhere: in high-stakes work, the output should expose its own uncertainty instead of laundering it into a confident single number.